Study On Visual Neural Network Computation Model And Its Applications In Image Processing

The visual system is the main way for human to perceive the objective world. With neuralexperimental means improving, it is possible to research the neural decoding and encoding patternof the visual system with the single cell, groups of visual function and the complex informationflow of visual cortex. But these researches mostly limited to observation and verification in visualneurophysiological experiments. Therefore, it is meaningful to apply the model of visual neuralcharacteristics to image processing in artificial intelligence field and information processing field.Firstly, this paper used neuronal receptive field and fluctuation characteristics of neuronal firing rateto explain the important role of the function of the photosensitive layer when getting theinformation; Secondly, this paper discussed the function of the neural encoding based on thechemical synaptic plasticity in signaling processes to explain the importance of neural encodingpattern when the visual system extracts the image information and applies it to image edgedetection; Finally, computed tomography image was used for detecting the edge of parts’ defectwith the combination of the function of the photosensitive layer and the characteristics of neuralencoding based on the dynamic chemical synapses. The work and research results of this paper aresummarized as follows:(1) This paper proposed a new method of image edge detection based on the function ofphotoreceptor in visual system. Firstly, the neural network was constructed with LIF neuronalelectrophysiological model. Secondly, the each neuron would be classified as excitation (ON)type or inhibition (OFF) type according to the neural firing pattern. And then the weak edgeswere highlighted by using center-surround antagonistic receptive field feature and feedbackenhancing mode of neuronal excitation. Meanwhile the image movement in multi-direction andmulti-scale was applied to overcome the adaptability of photoreceptor and highlight thecontrast of weak details. Finally edge image was acquired by fusing the variance information,such as the photosensitive neural network’s firing rates. The result proved that the new methodcan effectively detect the intact multi-intensity edge image, especially the weak edge detectionis improved significantly.(2) A new method of image edge detection based on the neural encoding of dynamic chemicalsynaptic was proposed in this paper. Firstly, three neural networks were constructed with LIFneuronal electrophysiological model. Then the signal conversion capability of the chemicalsynaptic plasticity was adjusted based on the center(ON)-surround(OFF) receptive field featureto highlight the neural encoding of the spatial feature of image and separate neural firing patterns in edge and non-edge region; Image movement was applied to eliminate encoding innoise; The encoding was affected by the temporal characteristic of neural network firing so thatthis paper strengthened it based on the feedback from neural spike to chemical synaptic todetect edge more effectively. The result proved that the new method can effectively detect theedge, express the details fully.(3) This paper proposed a new method to detect the edge of defect in the parts and attempted toapply visual neural network calculation model to analyze the computed tomography(CT) imageof parts. Firstly, based on the function of photosensitive layer neural network and chemicalsynaptic plasticity, the ability of signal conversion of chemical synapse was adjusted with thefluctuation characteristics of neural firing rate, and the encoding of the visual perceptioncharacteristics of LIF neuron network was used to detect the edge of parts’ defect in the CTimage. The results proved that compared to traditional methods, the new method can detect thedetails greatly and have great realizability.